OpenAI has added a new large-scale language model (LLM) to ChatGPT and its API, called GPT-4o mini. As the name suggests, the GPT-4o Mini model is a smaller version of the GPT-4o model introduced in May. The mini model is designed to balance the power of GPT-4o with a more cost-efficient approach.
GPT-4o mini has many of the features of its larger sibling, but the API currently only supports text and vision, with image, video, and audio input and output still in development. Like GPT-4o, the new model has a context window of 128,000 tokens, eight times larger than GPT-3.5 Turbo. The new model also comes with enhanced safety features. In addition to the features already built into GPT-4o, GPT-4o mini adds new techniques to make it more resistant to jailbreaking, improper prompt injection, and other issues that developers face when trying to widely deploy AI APIs.
Preparing for bigger jobs
OpenAI suggests that the larger context window and other upgrades, such as improved understanding of non-English text, will make GPT-4o mini especially useful for processing large documents and linking multiple interactions with AI models. For example, it could provide better recommendations in online stores, speed up real-time text responses in customer service, and give students studying for exams faster, more accurate, and more detailed answers than other models. OpenAI has a vision for GPT-4o to automate and streamline business processes with its ability to take data and take actions in external systems. For companies using the API, costs will be significantly reduced to just over half the price per token of GPT-3.5 Turbo.
“OpenAI is committed to making intelligence as widely accessible as possible,” OpenAI explained in the announcement. “By making intelligence much more affordable, we expect GPT-4o mini will greatly expand the range of applications built with AI.”
GPT-4o mini is part of a recent wave of small-scale LLMs like Google's Gemini Flash and Anthropic's Claude Haiku. But according to OpenAI, GPT-4o mini blows them away when it comes to many standard tests. The model scored 82% on the Massive Multitask Language Understanding (MMLU) benchmark, compared to 77.9% and 73.8% for Gemini Flash and Haiku, respectively. The same goes for the MGSM and Human Eval tests, where GPT-4o Mini achieved 87% and 87.2%, compared to 75.5% and 71.5% for Gemini Flash and 71.7% and 75.9% for Haiku. In other words, as shown in the chart below, GPT-4o Mini wins in text understanding in addition to math and coding tasks.
Mini Model Main Plan
According to OpenAI, the introduction of GPT-4o Mini is an important step towards making advanced AI more affordable and accessible. The reduced cost and improved performance will make it easier to incorporate AI into everyday applications. ChatGPT users will have access to the models starting this week. OpenAI also plans to introduce fine-tuning capabilities for GPT-4o Mini within its API.
Looking at the bigger picture, it marks another step forward for ChatGPT's evolving service. As OpenAI phases out GPT-3.5 for ChatGPT, the focus shifts to the next stage of providing a more powerful model. OpenAI CEO Sam Altman has long hinted at how GPT-5 will be a “significant improvement” over existing models. At the same time, the leak of OpenAI's scale for measuring AI power shows that we still have a long way to go before we can achieve the still-mythical artificial general intelligence (AGI) that can perfectly mimic how the human mind works.
